Remember when Meta’s AI assistant quietly showed up in the search bar of your apps, sitting there like a new roommate nobody introduced you to? You could ask it things. Most people poked at it once, decided it was a chatbot with a blue swoosh, and went back to scrolling. That version of Meta’s AI was a talker. The next one, apparently, is a doer.
Meta plans to launch an AI agent platform called Hatch in the coming weeks. It’s aimed squarely at what OpenAI and Anthropic are building, it will handle tasks based on what you tell it to do, and it’s expected to start behind a waitlist. One more detail makes it different from the rest of the field: Hatch will use social data from Instagram and Facebook to do its work.
Doer versus talker, in plain terms
If you’ve read anything on this site before, you know I keep coming back to one distinction, because it’s the whole ballgame.
- A chatbot answers. You ask what to cook, it gives you a recipe.
- An agent acts. You say “handle dinner,” and it goes off and does the steps.
An agent takes an instruction and breaks it into a sequence of actions, then carries them out without you approving each one. That’s the appeal, and that’s also the part that makes people uneasy. Answers you can ignore. Actions have consequences.
So when Meta says Hatch performs tasks based on user instructions, that’s the category it’s entering. Not “smarter search.” Something closer to delegation.
The social data wrinkle
This is the piece I find genuinely interesting, and I want to be careful about it, because it cuts both ways.
Most agents you’ve heard about start cold. You open one up, and it knows nothing about you beyond what you type into the box. You spend the first ten minutes explaining your preferences, your budget, your general vibe. Every session, a little bit of re-explaining.
Hatch is designed to pull from Instagram and Facebook data. That means an agent that arrives with context: the accounts you follow, the things you’ve engaged with, the shape of your interests as expressed over years of tapping around two apps.
For the person using it, that could feel remarkable. Less setup. Fewer clarifying questions. An assistant that already has a sense of what you like without you narrating it.
For the same person, thinking about it for another five seconds, it could feel like a lot. Your social history was collected in a context where it fed a recommendation feed. Using it to drive actions on your behalf is a different job for the same pile of information. Not automatically worse, but different, and worth understanding before you say yes.
Why the waitlist matters more than it sounds
A waitlist looks like a marketing move, and partly it is: scarcity gets attention. But there’s a practical reason agent products keep launching this way.
Agents fail in ways chatbots don’t. A chatbot that misunderstands you produces an awkward paragraph. An agent that misunderstands you produces an action, and actions ripple outward into other systems and other people’s inboxes. Rolling that out gradually is how a company finds the sharp edges without finding them in front of billions of users at once.
So if you’re on a waitlist for something like this, you’re not just in a queue. You’re part of the testing.
What I’d actually watch for
When Hatch shows up, the marketing will tell you what it can do. Here’s what I’d want to know instead, and these are fair questions to ask of any agent product, not just this one.
- What exactly can it act on? There’s a big gap between drafting something for your approval and completing something on its own.
- How do I see what it did? A clear log of actions taken is the difference between a tool and a black box.
- Can I turn the data part off? If the social-data connection is what makes it useful, is there a version that works without it, and how much worse is that version?
- Where does it stop and ask? Every agent needs a line where it pauses for a human. Knowing where that line sits tells you how much trust you’re extending.
The shift underneath the announcement
Meta competing with OpenAI and Anthropic on agents is the headline. The more interesting story is what each company brings to the fight. OpenAI and Anthropic bring models and developer ecosystems. Meta brings something nobody else has in the same volume: a long, detailed record of what a couple billion people find interesting.
That’s a real advantage for building an agent that feels personal from the first minute. It’s also the reason this launch deserves a slower read than most. The question isn’t whether an agent that knows you can be useful. It obviously can. The question is whether you’re comfortable with the version of you it’s working from, and whether you get any say in that.
We’ll know a lot more when it’s out. For now, my advice is the same as always: get clear on what you’re delegating before you delegate it.
đź•’ Published: